How Anthropic expects agent infrastructure to change

ZUSAMMENFASSUNG

In 30 Sekunden

As models become better at deciding what steps to take, Anthropic’s Angela Jiang expects developers to remove some rigid workflow rules and combine agents in new ways, such as having one challenge another’s proposal. Jiang also describes agents requesting specific permissions and acting through auditable identities as a possible operating model. Jiang recommends measuring value first in individual speed, then team productivity and cross-company processes. She also warns that cheap individual prototypes can produce organizational sprawl when nobody coordinates them.

1An agent may ask for specific access before starting workJiang envisions an agent requesting the resources needed for an outcome, receiving only approved access and operating through an auditable identity. She describes this as an emerging operating model.
2Better models may need fewer rigid workflow rulesJiang says elaborate step-by-step frameworks can be fragile. She expects thinner frameworks and combinations such as competing agents or one agent challenging another.
3A factory agent can preserve knowledge that leaves with an expertJiang describes a hackathon project that combined operating procedures and monitoring signals to approximate an experienced employee’s judgment about machinery.
4Security, compliance and evaluation limit adoptionLesse and Jiang say organizations need appropriate guardrails and ways to evaluate agents before they can use them more broadly.
5Start measuring agent value with individual speedJiang suggests moving from an individual’s speed to team productivity and then cross-company processes, before adding financial and user metrics.
6Cheap prototypes can create organizational sprawlJiang says giving everyone the ability to build independently can make coordinated, coherent work harder without a shared direction.
7A desired outcome and budget could become the interfaceLesse describes a possible future in which a person sets what good work looks like and a spending limit, and an agent iterates toward that result.

Das Wichtigste

Give agents clear outcomes and guardrails rather than assuming every step must be hard-coded, Jiang suggests.
Consider auditable identities and explicit permissions for the resources an agent needs.
Measure gains at the individual and team level before expanding to cross-company workflows, Jiang recommends.
Coordinate independently built prototypes around a shared direction to limit duplicated or fragmented work.
Security, compliance and evaluations remain barriers organizations need to address before wider adoption.

ThemenAgent Engineering Agenten-Authentifizierung Multi-agent systems Harness engineering KI-Engineering-Führung

VOLLTEXT

Den bearbeiteten Text nach Kapiteln lesen

Transkript zeigen

Nur das Transkript

From model access to managed agent workflows

Anthropic’s platform team describes a move from model inference toward managed infrastructure. Jiang discusses how an agent identity might request and receive limited permissions.

Jess Yann says the Claude Platform has moved beyond an API for model inference toward features that take some infrastructure and agent-framework work off a team’s hands. Asked about customer feedback, Angela Jiang points to managed-agent features such as memory, outcomes and “dreaming,” as well as the level of abstraction offered to developers.

Jiang says agent identity and permissions are still evolving. She imagines an agent receiving an intended outcome, asking for access to the resources needed and proceeding with only what the user approves. An auditable service identity could make it possible to inspect whether the agent is doing the intended work. She presents this as a likely future operating model, not a finished design.

Agents communicate through exposed interfaces

Lesse says teams have exposed agent capabilities through interfaces that other agents can call. Better models and longer-running infrastructure make more of these workflows possible.

Katelyn Lesse says one agent can expose an API or another interface so a second agent can interact with it. She has seen teams build a thin MCP server around an agent, allowing another agent to call it. These interfaces make it possible to combine capabilities rather than have every agent perform every task itself.

Lesse attributes the viability of such workflows partly to better models. Previously, developers needed more scaffolding and fixed procedures to keep unpredictable behavior on track. She says newer models can identify steps within reasonable guardrails, and improved infrastructure lets agents run longer, respond to a workspace event and return when their work is ready.

Thinner frameworks and teams of agents

Jiang expects fewer rigid rules around increasingly capable models. She describes agents competing, challenging one another or asking an adviser for help.

Jiang says earlier agent systems often encoded complicated business processes as rigid sequences: step A could lead to B only under specific conditions. Those structures could be fragile and constrain what the model was able to do. As models improve at reasoning and using tools, she expects developers to remove some restrictive rules and make the framework around the model thinner.

That does not mean abandoning all structure. Jiang describes higher-level strategies in which several agents try to solve the same problem, one agent challenges another’s proposal, or a stuck agent asks a more capable adviser for help. Yann suggests combining approaches—for example, selecting among several attempts and then iterating on the chosen one.

Preserving factory knowledge and extending development

A hackathon project combined procedures and factory monitoring to retain expertise. Lesse also describes development agents working from requirements through testing.

Jiang recalls a hackathon project for manufacturing facilities where maintaining machinery can depend on one expert’s decades of experience. The builder combined operating procedures, equipment monitoring and signals from the plant so agents could approximate some of that judgment. In her account, the system offered a way to retain and reuse knowledge when an experienced employee retired.

Lesse sees development teams applying agents beyond writing code. They are considering the entire project: defining requirements, running a development environment, testing the resulting code and carrying out quality assurance. She cites Shopify’s work on a system called River as one example of more customized, end-to-end development workflows.

Adoption depends on safeguards and evaluation

The panel identifies security, compliance and evaluations as practical barriers. Jiang says organizations must revisit assumptions and checklists as agents change how work is done.

Asked why not every organization is seeing the same gains, Lesse names security and compliance safeguards and the ability to evaluate agents. Jiang agrees that teams must revisit security assumptions and checklists built for earlier ways of working. Organizations need confidence that the agents are operating within appropriate boundaries.

Measure value from individuals to teams

Jiang recommends starting with individual speed, then team productivity and finally cross-company processes. Lesse describes engineers coordinating agents while retaining responsibility for systems.

For measuring return on investment, Jiang recommends starting smaller than a company-wide list of processes to automate. First ask how much faster an individual can work, then look at team productivity. Only after that, she suggests, should an organization tackle processes spanning multiple teams with different expertise and operating procedures.

Speed and productivity can be early measures, Jiang says; as adoption progresses, companies can add financial and user metrics tied to the outcomes they care about. The staged approach is intended to reveal where value appears as individual work connects to team and company workflows.

Lesse says her engineering teams still need people who understand the systems they build, know how to operate them and can respond when something goes wrong. What has changed is how those people work: rather than a single technical lead directing a group that picks up tickets, more engineers can shape end-to-end design and coordinate agents to carry out work.

Avoid sprawl and imagine a less visible agent layer

Jiang warns that cheap independent prototypes can make coordination harder. She envisions agents embedded in familiar workflows and adapting to team preferences.

Jiang sees a potential failure mode in the new independence. When creating a prototype is cheap, one person can build ten options instead of choosing one beforehand. That can be useful, but without a shared direction the result may be sprawl, and it may become harder to coordinate the pieces into a coherent system.

Looking ahead, Jiang imagines agents embedded in familiar organizational interfaces rather than a collection of separate tools people deliberately open. An agent might notice a problem, investigate it and bring a proposed code change for review. With appropriate permission, a user might later allow it to handle small recurring fixes without asking each time.

In that vision, a workflow shared by two people could have an agent that understands their preferences; a different team’s agent could reflect a different set. Jiang compares the resulting common layer to an operating system that fills gaps in team workflows, while keeping familiar interfaces.

Specify an outcome and a budget

Lesse describes the platform’s outcome rubric and a possible future interface where a person provides a goal and spending limit. The agent would iterate toward the result.

Lesse points to the platform’s outcomes feature as a step toward that direction: a person specifies what good work looks like in a rubric and how many attempts the agent may make. She imagines a future in which someone gives an agent an outcome and a budget—for example, to summarize interview notes and prepare feedback—and lets it work toward the result without managing every step.